Improved Performance when using Window.partitionBy instead of groupBy

I have one scala-spark dataframe with three columns say Col1, Col2 and Col3. This dataframe has around 54 million rows.

What I want is max of Col3 per Col1,Col2 combination like below

[

val maxDf = df.groupBy(Col1, Col2).agg(F.max(Col3))

[

But when I am trying to achieve above using groupBy somehow it is failing with below error

java.lang.OutOfMemoryError: Unable to acquire 8168736160 bytes of memory, got 0

Then I tried the same using Window function

val window  = Window.partitionBy("Col1","Col2")

val maxDf = df.withColumn("maxVal", F.max("Col3").over(window))

Somehow above code got executed successfully. I am trying to find out the reason behind this as theoretically both the operation are similar and have same execution plan.